The neural network model section only has 20+ lines of code!
但却可以识别手写体的0-9这十个数字！
But it can identify the handwriting 0-9 of these ten figures!
训练所用的手写体样本：
Handbook for handwriting used for training:
Code: Select all
For Idx = 0 To #MAI_Sample-1
ForEach _DimSample(Idx)\ListSample()
For r = 0 To #MAI_Sample-1
;对样本像素量化值进行卷积
Convolue.f = _DimWeight(r, 0) * -1
For c = 1 To #MAI_Factor
Convolue + _DimWeight(r, c) * (_DimSample(Idx)\ListSample()\Pixels[c-1])
Next
;计算出可能性
OutResult.f = 1/(1+Exp(-Convolue))
;对预期值进行比对,
RFeedback.f = Exp(-Convolue)/Sqr(1+Exp(-Convolue))
If Idx = r And OutResult < #MAI_Judged
AntiConvolue.f = (0.90 - OutResult) * RFeedback
ElseIf Idx <> r And OutResult >= #MAI_Judged
AntiConvolue.f = (0.10 - OutResult) * RFeedback
Else
Continue
EndIf
;如果达不到预期效果,对权限矩阵进行微调,以达成在下一次计算时,更加接近预期值!
_DimWeight(r, 0) - 0.01 * AntiConvolue
For c = 1 To #MAI_Factor
_DimWeight(r, c) + 0.01 * AntiConvolue * _DimSample(Idx)\ListSample()\Pixels[c-1]
Next
Next
Next
Next
;只训练一次的效果 <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<
;没有一个识别成功
; 第 1次训练
; : 0 , 1 , 2 , 3 , 4 , 5 , 6 , 7 , 8 , 9 ,
; 0: -, -, -, -, -, -, -, -, -, 51.58%,
; 1: -, -, -, -, -, -, -, -, -, 51.35%,
; 2: -, -, -, -, -, -, -, -, -, 50.23%,
; 3: -, -, -, -, -, -, -, -, -, 51.78%,
; 4: -, -, -, -, -, -, -, -, -, 53.33%,
; 5: -, -, -, -, -, -, -, -, -, 51.70%,
; 6: -, -, -, -, -, -, -, -, -, 51.07%,
; 7: -, -, -, -, -, -, -, -, -, 53.54%,
; 8: -, -, -, -, -, -, -, -, -, 53.39%,
; 9: -, -, -, -, -, -, -, -, -, 52.98%,
Training 50 times the effect：
;训练50次的效果 <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<
;有明显的进步了,样本识别方面,0,2,3,4,5,6,7已经能够成功识别
; : 0 , 1 , 2 , 3 , 4 , 5 , 6 , 7 , 8 , 9 ,
; 0: 51.25%, -, -, -, -, -, -, -, -, -,
; 1: -, -, -, -, -, -, -, -, 50.54%, -,
; 2: -, -, 52.02%, -, -, -, -, -, 53.32%, -,
; 3: -, -, -, 59.98%, -, -, -, -, -, 60.29%,
; 4: -, -, -, -, 55.66%, -, -, -, -, -,
; 5: -, -, -, -, -, 56.79%, -, -, -, -,
; 6: -, -, -, -, -, -, 53.67%, -, -, -,
; 7: -, -, -, -, -, -, -, 53.67%, -, 54.85%,
; 8: -, -, -, -, -, -, -, -, 62.38%, -,
; 9: -, -, -, -, -, -, -, -, -, 69.44%,
Training 500 times the effect：
;训练500次的效果 <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<
;样本识别方面,已经没问题了,目标识别,还有几个不理想,这个跟训练样品数量有限有很大的关系,
; : 0 , 1 , 2 , 3 , 4 , 5 , 6 , 7 , 8 , 9 ,
; 0: 58.03%, -, -, -, -, -, -, -, -, -,
; 1: -, 52.46%, -, -, -, -, -, -, -, -,
; 2: -, -, 65.46%, -, -, -, -, -, -, -,
; 3: -, -, -, 58.69%, -, -, -, -, -, -,
; 4: -, -, -, -, 57.11%, -, -, -, -, -,
; 5: -, -, -, -, -, 57.51%, -, -, -, -,
; 6: -, -, -, -, -, -, 57.30%, -, -, -,
; 7: -, -, -, -, -, -, -, 55.08%, -, -,
; 8: -, -, -, -, -, -, -, -, 66.43%, -,
; 9: -, -, -, -, -, -, -, -, -, 58.93%,
完整的代码：
Complete code:
Code: Select all
; **************************************************
; ***** 手写数字智能识别简例--单层神经网络 *****
; ***** 迷路仟 QQ:714095563 *****
; ***** 2017.04.27 *****
; **************************************************
; 说明:
; 1.本实例是用于识别手写体数字的神经网络代码,旨在教学与案例,
; 对神经网络作一个简约的编码. 本实例采样数量有限,所以,在识别手写体数字,并非万能,
; 如果有兴趣的同行,可以给程序喂食更多的手写体样本,以达到更加智能的效果!
; 2.本实例采用的是单层神经网络算法,在识别方面,不如多层神经网络的效果更好!
; 3.本实例没有对样本有识别目标做滤波处理,故本实例没有什么实战意义,
; 只用于阐述神经的基本结构!
#MAI_LightR = 299
#MAI_LightG = 587
#MAI_LightB = 114
#MAI_Total = #MAI_LightR + #MAI_LightG + #MAI_LightB
#MAI_Factor = 28*28
#MAI_Sample = 10
#MAI_Judged = 0.5
; 字符转换结构
Structure __FactorInfo
Pixels.f[28*28]
EndStructure
; 字符表
Structure __SampleInfo
List ListSample.__FactorInfo()
EndStructure
UsePNGImageDecoder()
Global Dim _DimSample.__SampleInfo(#MAI_Sample-1)
Global NewList _ListIdentify.__FactorInfo()
Global Dim _DimWeight.f(#MAI_Sample-1, #MAI_Factor) ;为每个数字设置一个权限矩阵
; 样本图片采样
Procedure MAI_LoadSampleImage(ImageName$)
;20x10数字阵,像素大小28x28
ImageID = LoadImage(#PB_Any, ImageName$)
If ImageID
CreateImage(0, ImageWidth(ImageID)+20, ImageHeight(ImageID)+20) ;为采样进行偏移处理做装备
If StartDrawing(ImageOutput(0))
Box(0, 0, 1000,1000, 0)
DrawImage(ImageID(ImageID), 10, 10)
For r = 0 To 9
For c = 0 To 19
For k = 1 To 5
i = Random(5,0)-3 ;这里对图片采样进行偏移采样,以获取更多可能性
j = Random(5,0)-3
;==============
AddElement(_DimSample(r)\ListSample())
Idx = 0
For y = 0 To 27
For x = 0 To 27
Color = Point(10+c*28+i+x, 10+r*28+j+y)
;通过灰度值,将样本像素化为1或0
If ((Red(Color)*#MAI_LightR+Green(Color)*#MAI_LightG+Blue(Color)*#MAI_LightB)/#MAI_Total) & $FF >= 128
_DimSample(r)\ListSample()\Pixels[Idx] = 1
Else
_DimSample(r)\ListSample()\Pixels[Idx] = 0
EndIf
Idx+1
Next
Next
Next
;==============
Next
Next
StopDrawing()
EndIf
FreeImage(ImageID)
EndIf
EndProcedure
;-
; 训练样本
Procedure MAI_TrainSampleImage(CountTrain)
Debug "初始化"
For r = 0 To #MAI_Sample-1
For c = 0 To #MAI_Factor
_DimWeight(r,c) = Random(2,1)/#MAI_Factor/#MAI_Sample
Next
Next
For k = 1 To CountTrain
Debug "第"+RSet(Str(k), 4, " ")+"次训练"
For Idx = 0 To #MAI_Sample-1
ForEach _DimSample(Idx)\ListSample()
For r = 0 To #MAI_Sample-1
;对样本像素量化值进行卷积
Convolue.f = _DimWeight(r, 0) * -1
For c = 1 To #MAI_Factor
Convolue + _DimWeight(r, c) * (_DimSample(Idx)\ListSample()\Pixels[c-1])
Next
;计算出可能性
OutResult.f = 1/(1+Exp(-Convolue))
;对预期值进行比对,
RFeedback.f = Exp(-Convolue)/Sqr(1+Exp(-Convolue))
If Idx = r And OutResult < #MAI_Judged
AntiConvolue.f = (0.90 - OutResult) * RFeedback
ElseIf Idx <> r And OutResult >= #MAI_Judged
AntiConvolue.f = (0.10 - OutResult) * RFeedback
Else
Continue
EndIf
;如果达不到预期效果,对权限矩阵进行微调,以达成在下一次计算时,更加接近预期值!
_DimWeight(r, 0) - 0.01 * AntiConvolue
For c = 1 To #MAI_Factor
_DimWeight(r, c) + 0.01 * AntiConvolue * _DimSample(Idx)\ListSample()\Pixels[c-1]
Next
Next
Next
Next
Next
EndProcedure
;-
; 随机识别样本,以判断学习进度
Procedure MAI_IdentifySample()
Text$ = " : "
For Idx = 0 To #MAI_Sample-1
Text$ + RSet(Str(Idx)+" ", 6, " ") +", "
Next
Debug Text$
For Idx = 0 To #MAI_Sample-1
Index = Random(19,0)
SelectElement(_DimSample(Idx)\ListSample(), Index)
*pFactor.__FactorInfo = _DimSample(Idx)\ListSample()
Text$ = Str(Idx)+": "
For r = 0 To #MAI_Sample-1
Convolue.f = _DimWeight(r, 0) * -1
For c = 1 To #MAI_Factor
Convolue + _DimWeight(r, c) * (*pFactor\Pixels[c-1])
Next
OutResult.f = 1/(1+Exp(-Convolue))
If OutResult >= #MAI_Judged
Text$ + RSet(StrF(OutResult*100, 2)+"%", 6, " ") +", "
Else
Text$ + RSet("-", 6, " ") +", "
EndIf
Next
Debug Text$
Next
Debug "====="
EndProcedure
MAI_LoadSampleImage("训练样本_数字.png")
MAI_TrainSampleImage(10)
MAI_IdentifySample()
